Enhancing CNN Performance Through Tiny Image Sets and Incremental Learning Techniques

Artificial Intelligence (AI)-based solutions for visual recognition are increasingly used in applications aiming to monitor the emergence of environmental phenomena, such as wildfire. Mostly, the solutions envisioned aim to autonomously identify the insurgence of an event from the prompt recognition of the first consequences: fire, smoke, and/or abnormal high-temperature, to mention a few. These techniques leverage the Edge computing attitude to identify the event by exploiting the Convolutional Neural Networks (CNNs) to analyze perceived images. Especially in the case of wildfire recognition, the environment monitored by the model is slowly but continuously subjected to modification, and moreover, the sensing elements may degrade their performance due to low maintenance, accidental events, or device degradation over time.Techniques such as Incremental Learning (IL) are meant to enable Artificial Intelligence (AI) models to continuously learn, adapting their behavior to the evolving conditions of the analyzed scenario. The learning rate of IL-based models characterizes the degree of adaptability and may require several cycling training phases to adjust to evolving conditions. Furthermore, Catastrophic Forgetting (CF) events can affect these category models, degrading their performances (e.g., accuracy level).This paper presents a solution that leverages the continuous learning workflow driven by a distributed approach that generalizes the knowledge acquired by the model and minimizes the issues produced by CF events. This goal is achieved by exploiting the DILoCC solution, which is extended to support the learning distribution among Internet of Things (IoT) devices.